Feature selection on feed-forward neural network for Eurex PSE-MSCI index futures

The stock index provides a natural benchmark for the stock market performance. More importantly, the stock index is the underlying of the futures. The paper focused on predicting the Eurex PSE-MSCI index futures price by maximizing an artificial neural network (ANN). Specifically, the study created a financial tool using the feed-forward network (FFNN) principles that forecast index futures price using technical, economic, and the aggregated indicators as the features. Moreover, extreme gradient boosting (XGBoost), an ensemble method under the embedded feature selection (FS), was employed. The FS application effectively removed the redundancy and noise caused by uncorrelated variables. More importantly, the study found that the neural network can provide an accurate forecast of the futures price during regular market periods but showed less accuracy when presented with market surprises. Although all indicators provide an accurate forecast with less than 5% mean absolute percentage error, economic indicators have the largest deviation and lowest accuracy among all indicators. Meanwhile, technical and aggregated indicators have the same accuracy level, with the unfiltered technical indicators showing the highest accuracy. Although it had a noticeable improvement in the forecasts, extreme gradient boosting showed no significant improvement in the neural network\u27s predictability